Noise Reduction and Troubleshooting in Neural Recording
Identify and remove environmental, system and biological noise in neural recordings, with a systematic troubleshooting sequence.
8 min read
Updated October 12, 2026
Noise in a neural recording almost always comes from one of three places: the environment, the recording chain, or the animal. Identify which one before changing anything on the rig. Changing several things at once is how a solvable problem becomes a permanent one.
Research use only. NeuroNexus products and the procedures described here are for non-clinical research use, and must not be used in human or veterinary medical procedures.
Where noise comes from#
Environmental#
| Source | Frequency | Pattern | Effect on the recording |
|---|---|---|---|
| Mains | 50/60 Hz and harmonics | Sinusoidal, stable | Obscures low-frequency oscillations |
| Broadband EMI | Wide | Variable, equipment-dependent | Raises the floor across bands |
| RF | MHz–GHz | Intermittent | Broadband floor, often rectified into the signal band |
| Fluorescent lighting | 50/60 Hz + harmonics | Periodic switching | Baseline noise and transients |
The usual culprits are power strips and supplies, monitors and computers, and wireless devices. Building infrastructure adds HVAC, elevators and electrical panels.
The recording chain#
| Component | Noise | Frequency | Mitigation |
|---|---|---|---|
| Headstage / amplifier | Thermal, shot | Broadband | Low-noise front end; digitize early |
| Cables | EMI pickup | Variable | Shielded, twisted pair; short pre-amp run |
| Connectors | Contact resistance | DC and low | Clean, secure, strain-relieved |
| Power supplies | Switching | kHz | Linear supplies or filtering |
Three configuration faults account for most of what is blamed on components:
Ground loops. Two ground paths enclose an area, and that area is an antenna for magnetic fields.
Impedance mismatch. A high-impedance site raises thermal noise and forms a divider with the amplifier input. Both reduce signal-to-noise before any amplification happens.
Incomplete shielding. A cage with an ungrounded panel, or a cable routed alongside a power line, defeats the rest of the installation.
Biological#
| Artifact | Frequency | Amplitude | Source |
|---|---|---|---|
| EMG | 10–500 Hz | 10–1,000 µV | Facial, neck and limb muscle |
| ECG | 0.5–40 Hz | 10–100 µV | Cardiac activity |
| EOG | 0.1–10 Hz | 50–500 µV | Saccades and blinks |
| Respiration | 0.1–2 Hz | Variable | Breathing movement |
Movement contributes three more: triboelectric noise from cable motion, micro-motion at the tissue interface, and gross movement. The last appears on every channel at once.
Reducing environmental noise#
Faraday cage. Copper mesh or solid aluminium, grounded at a single point, with filtered feed-throughs for cables, and large enough to enclose the recording setup. A properly installed cage takes 20–40 dB off mains interference. A cage with an ungrounded door takes off nothing.
Distance and scheduling. Keep supplies, monitors and non-essential electronics one to two meters away. Use a dedicated circuit. Turn off what you do not need during the recording.
Battery power for headstages, pre-amplifiers and stimulators removes the mains coupling path. Computers and monitors stay on mains.
Power conditioning. Isolation transformers, a pure-sine UPS, and line filters on whatever remains on mains.
Grounding and referencing#
Ground at one point. Every chassis, the cage, and the subject ground go to the same point, over a low-impedance connection. More than one ground path is a loop.
Reference placement decides what you subtract. Put the reference in stable, electrically quiet tissue, well away from stimulation sites, with a large surface area so its own potential is stable. A reference in active tissue subtracts real signal.
Keep impedance ↗ in range. Extracellular sites in roughly 0.1–2 MΩ; EEG and ECoG contacts in the kilohm range. Measure before each session, track the values, and treat a rising trend as a failing site rather than a bad reading.
Optimizing the recording chain#
Digitize early. The canonical layout is electrode → short cable → low-noise headstage → long cable → acquisition system. Data after the headstage is digital, which is why cable movement stops mattering once the signal has been converted.
Cables. Twisted pair for differential signals, shield grounded at one end only, as short as the rig allows, routed away from power.
Gain and bandwidth. Set gain to use the converter range without clipping, and match bandwidth to the signal. Over-amplification does not add information and costs headroom.
Sampling. At least 2.5× the highest frequency of interest, with an anti-alias filter ahead of the converter.
Reducing biological noise#
Acute. Stable stereotaxic fixation, appropriate anesthetic depth, maintained physiology, and minimal handling during recording.
Chronic. Allow full recovery before recording, manage the cable so nothing tugs the implant, and habituate the animal to the tether before the data sessions begin.
Placement. Where the anatomy allows, keep sites away from major muscle groups. EMG sits in the same band as gamma, and no filter separates them afterwards.
Filtering — and its limits#
| Filter | Use | Range | Caution |
|---|---|---|---|
| High-pass | Remove drift and movement | > 0.1–1 Hz | Above 300 Hz for spike-band work only |
| Low-pass | Anti-aliasing | < 5–10 kHz | Match to sampling rate |
| Notch | Mains interference | 50/60 Hz | Removes neural activity in the same band |
| Band-pass | Isolate a band | Variable | Defensible only if the band is defined in advance |
A notch filter is a last resort. It removes the interference and the neural activity underneath it, and it leaves the ringing at the band edges in your data. Fix the grounding instead.
Where the artifact is deterministic — stimulation, for instance — template subtraction removes it without touching the rest of the band. Common average referencing removes what is common to all channels, and is effective when the noise is genuinely common-mode.
A diagnostic sequence#
Work in order. Each step rules out a class of cause. Skipping ahead means changing something you have not shown to be at fault.
1. Look. Cable connections secure, no visible electrode damage, ground connected, power on.
2. Measure impedance. All recording channels, plus reference and ground. Compare against the baseline you recorded at implant.
3. Disconnect from the subject and record. This separates system noise from biological noise in one measurement, and it is the step most often skipped.
4. Assess the environment. Record inside and outside the cage. Turn off nearby equipment one at a time, not all at once.
5. Test the signal path. Swap cables, then headstages. Check connector condition. Verify gain and filter settings are what you think they are.
6. Check grounding. Confirm single-point grounding, test alternative reference positions, and look for loops with a multimeter.
7. Go to the frequency domain. Record ten-minute segments under different conditions, compute power spectral density, and identify the offending frequencies. Compare across channels: noise present on every channel is common-mode; noise on one is that channel.
Reading the result#
The same steps as a triage. Work down the table. The first test that fails locates the fault.
| Test | Failing result | Where the fault is |
|---|---|---|
| Impedance on the affected channels | Outside the implant baseline | The electrode or its connections — clean, reseat, or replace |
| Power spectral density at 50/60 Hz | Mains peak present | Grounding and shielding |
| Artifact band and timing | An EMG, ECG, EOG or respiration signature | The subject — stabilization and site placement |
| Recording with the subject disconnected | Noise persists | The system — cables, headstage, supplies |
| All four above | Nothing isolated | Compare spectra across channels; common to all is common-mode |
Baseline and quality control#
Establish a baseline before you need it. Record five to ten minutes with the system connected but no subject. Measure RMS noise per channel and store the power spectral density. Every later diagnosis is a comparison against this.
Reasonable acceptance criteria for extracellular work: RMS noise below about 10 µV, no prominent mains peak, consistent levels across channels, no intermittent transients.
Re-measure impedance before each session, after any change to the rig, weekly for chronic implants, and after any troubleshooting.
Watch the recording as it runs. Waveform shape, amplitudes in the expected range, consistency across channels. Note artifact sources at the time; reconstructing them from the data afterwards is much harder.
Support#
Technical questions: support@neuronexus.com · +1.734.913.8858